In 45 out of 100 tries, AI writes code with a security hole in it. It compiles and the tests pass, so you never notice. Ask it for safe code, and scan what it writes.